v2.19.10

  - Increase rewards MA from 24 to 36 epochs.
  - Reset rewards for the given UID if the hotkey was changed.
  - Remove reliable UID sampling by default for API.

Author: dbobrenko

gpu_container/docker-compose.yml

@@ -2,7 +2,7 @@ services:
   gpu-app:
     build: .
     ports:
-      - "8000:8000"
+      - "127.0.0.1:8000:8000"
     environment:
       - MODEL_ID=WhereIsAI/UAE-Large-V1
       - VLLM_MODEL_ID=mrfakename/mistral-small-3.1-24b-instruct-2503-hf

prompting/weight_setting/weight_setter.py

@@ -1,5 +1,4 @@
 import asyncio
-import datetime
 import json
 from collections import deque
 from pathlib import Path
@@ -89,15 +88,17 @@ class WeightSetter(AsyncLoopRunner):
     """The weight setter looks at RewardEvents in the reward_events queue and sets the weights of the miners accordingly."""
 
     sync: bool = True
-    interval: int = 60 * 21
+    interval: int = 60 * 23
     reward_events: list[list[WeightedRewardEvent]] | None = None
     weight_dict: dict[int, list[float]] | None = None
     weight_syncer: WeightSynchronizer | None = None
 
+    # Rewards moving average persistency.
     reward_history_path: Path = Path("validator_rewards.jsonl")
-    reward_history_len: int = 24
-    # List of uids info per each epoch, e.g.: [{1: {"reward": 1.0}, 2: {"reward": 3.0}}].
-    reward_history: deque[dict[int, dict[str, Any]]] | None = None
+    # Rewards moving average, 36 epochs = approx. 12 hours.
+    reward_history_len: int = 36
+    # List of uids info per epoch, e.g.: [{1: {"reward": 1.0, "hotkey": "ABC"}, 2: {"reward": 3.0, "hotkey": "XYZ"}}].
+    reward_history: deque[dict[int, dict[str, float | str]]] | None = None
 
     class Config:
         arbitrary_types_allowed = True
@@ -120,17 +121,24 @@ async def _compute_avg_reward(self) -> npt.NDArray[np.float32]:
         num_uids = int(shared_settings.METAGRAPH.n.item())
         accum = np.zeros(num_uids, dtype=np.float32)
         if not isinstance(self.reward_history, deque) or len(self.reward_history) == 0:
-            logger.warning(f"Empty rewards history, setting zero weights: {self.reward_history}")
+            logger.error(f"Empty rewards history, setting zero weights: {self.reward_history}")
             return accum
 
-        for snapshot in self.reward_history:
-            for uid_str, info in snapshot.items():
-                accum[int(uid_str)] += float(info["reward"])
+        # Get current active hotkeys for the current set of UIDs.
+        active_hotkeys: dict[int, str] = {}
+        for uid_str, info in self.reward_history[-1].items():
+            active_hotkeys[int(uid_str)] = info.get("hotkey")
+
+        # Accumulate rewards for each epoch only if hotkey was not changed for the given UID.
+        for epoch_info in self.reward_history:
+            for uid_str, info in epoch_info.items():
+                if active_hotkeys[int(uid_str)] == info.get("hotkey"):
+                    accum[int(uid_str)] += float(info["reward"])
 
         avg = accum / len(self.reward_history)
         return avg
 
-    async def _save_rewards(self, rewards: npt.NDArray[np.float32]):
+    async def _update_rewards(self, rewards: npt.NDArray[np.float32]):
         """Persist the latest epoch rewards.
 
         The snapshot is appended to `reward_history` (bounded by `reward_average_len`) and the JSONL file at
@@ -142,20 +150,22 @@ async def _save_rewards(self, rewards: npt.NDArray[np.float32]):
         if not isinstance(self.reward_history, deque):
             self.reward_history = deque(maxlen=self.reward_history_len)
 
-        snapshot = {int(uid): {"reward": float(r)} for uid, r in enumerate(rewards)}
-        self.reward_history.append(snapshot)
+        hkeys = shared_settings.METAGRAPH.hotkeys
+        epoch_rewards: dict[int, dict[str, float | str]] = {}
+        for uid, reward in enumerate(rewards):
+            epoch_rewards[int(uid)] = {"reward": float(reward), "hotkey": hkeys[uid]}
+        self.reward_history.append(epoch_rewards)
 
-        tmp_path = self.reward_history_path.with_suffix(".jsonl.tmp")
-        block = getattr(shared_settings, "block", 0)
+        # block = getattr(shared_settings, "block", 0)
 
+        # Write results into tmp file, them move to the main rewards file to make write operation atomic.
+        tmp_path = self.reward_history_path.with_suffix(".jsonl.tmp")
         try:
             with tmp_path.open("w", encoding="utf-8") as file:
-                for snap in self.reward_history:
-                    row: dict[str, Any] = {
-                        "ts": datetime.datetime.now(datetime.timezone.utc).isoformat(timespec="seconds") + "Z",
-                        "block": block,
-                        "rewards": {str(k): v["reward"] for k, v in snap.items()},
-                    }
+                for epoch_rewards in self.reward_history:
+                    row: dict[str, Any] = {}
+                    for uid, info in epoch_rewards.items():
+                        row[str(uid)] = {"reward": float(info["reward"]), "hotkey": info["hotkey"]}
                     file.write(json.dumps(row, separators=(",", ":")) + "\n")
             tmp_path.replace(self.reward_history_path)
         except Exception as exc:
@@ -166,7 +176,7 @@ async def _load_rewards(self):
 
         Only the newest `reward_average_len` rows are retained.
         """
-        self.reward_history: deque[dict[int, dict[str, Any]]] | None = deque(maxlen=self.reward_history_len)
+        self.reward_history: deque[dict[int, dict[str, float | str]]] | None = deque(maxlen=self.reward_history_len)
         if not self.reward_history_path.exists():
             logger.info("No rewards file found - starting with empty history.")
             return
@@ -175,13 +185,16 @@ async def _load_rewards(self):
             with self.reward_history_path.open("r", encoding="utf-8") as file:
                 for line in file:
                     data = json.loads(line)
-                    payload = data.get("rewards")
-                    if payload is None:
+                    if not data:
                         raise ValueError(f"Malformed weight history file: {data}")
 
-                    self.reward_history.append({int(uid): {"reward": float(reward)} for uid, reward in payload.items()})
+                    epoch_rewards: dict[int, dict[str, float | str]] = {}
+                    for uid, info in data.items():
+                        epoch_rewards[int(uid)] = {"reward": float(info["reward"]), "hotkey": info.get("hotkey")}
+
+                    self.reward_history.append(epoch_rewards)
         except BaseException as exc:
-            self.reward_history: deque[dict[int, dict[str, Any]]] | None = deque(maxlen=self.reward_history_len)
+            self.reward_history: deque[dict[int, dict[str, float | str]]] | None = deque(maxlen=self.reward_history_len)
             logger.error(f"Couldn't load rewards from file, resetting weight history: {exc}")
 
     @classmethod
@@ -269,7 +282,7 @@ async def run_step(self):
                 logger.error("No rewards were found, skipping weight setting")
                 return
 
-            await self._save_rewards(final_rewards)
+            await self._update_rewards(final_rewards)
             averaged_rewards = await self._compute_avg_reward()
             averaged_rewards[averaged_rewards < 0] = 0
             averaged_rewards /= np.sum(averaged_rewards) + 1e-10

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "prompting"
-version = "2.19.9"
+version = "2.19.10"
 description = "Subnetwork 1 runs on Bittensor and is maintained by Macrocosmos. It's an effort to create decentralised AI"
 authors = ["Kalei Brady, Dmytro Bobrenko, Felix Quinque, Steffen Cruz, Richard Wardle"]
 readme = "README.md"

shared/settings.py

@@ -74,12 +74,12 @@ class SharedSettings(BaseSettings):
 
     SCORING_API_PORT: int = Field(8095, env="SCORING_API_PORT")
     # Hard-code MC validator axon, since it might be overwritten in the metagraph.
-    MC_VALIDATOR_HOTKEY: str = Field("5Cg5QgjMfRqBC6bh8X4PDbQi7UzVRn9eyWXsB8gkyfppFPPy", env="MC_VALIDATOR_HOTKEY")
-    MC_VALIDATOR_AXON: str = Field("184.105.5.17:42174", env="MC_VALIDATOR_AXON")
+    MC_VALIDATOR_HOTKEY: str = Field("5CGLCBndTR1BvQZzn429ckT8GyxduzyjMgt4K1UVTYa8gKfb", env="MC_VALIDATOR_HOTKEY")
+    MC_VALIDATOR_AXON: str = Field("184.105.215.196:42174", env="MC_VALIDATOR_AXON")
 
     # ==== API =====
     # Hotkey used to run api, defaults to Macrocosmos
-    API_HOTKEY: str = Field("5Cg5QgjMfRqBC6bh8X4PDbQi7UzVRn9eyWXsB8gkyfppFPPy", env="API_HOTKEY")
+    API_HOTKEY: str = Field("5CGLCBndTR1BvQZzn429ckT8GyxduzyjMgt4K1UVTYa8gKfb", env="API_HOTKEY")
     # Scoring request rate limit in seconds.
     SCORING_RATE_LIMIT_SEC: float = Field(5, env="SCORING_RATE_LIMIT_SEC")
     # Scoring queue threshold when rate-limit start to kick in, used to query validator API with scoring requests.

validator_api/chat_completion.py

@@ -237,7 +237,7 @@ async def chat_completion(
     uids: Optional[list[int]] = None,
     num_miners: int = 5,
     uid_tracker: UidTracker | None = None,
-    add_reliable_miners: int = 3,
+    add_reliable_miners: int = 0,
 ) -> tuple | StreamingResponse:
     # TODO: Add docstring.
     """Handle chat completion with multiple miners in parallel."""
@@ -252,7 +252,7 @@ async def chat_completion(
     primary_uids = filter_available_uids(
         task=body.get("task"), model=body.get("model"), test=shared_settings.API_TEST_MODE, n_miners=num_miners
     )
-    if uid_tracker is not None:
+    if uid_tracker is not None and add_reliable_miners > 0:
         # Add reliable uids, or ones with highest success rate to guarantee completed stream.
         reliable_uids = await uid_tracker.sample_reliable(
             task=TaskType.Inference, amount=add_reliable_miners, success_rate=0.99, add_random_extra=False